First-Order Diffusion Samplers Can Be Fast
Published:Dec 31, 2025 15:35
•1 min read
•ArXiv
Analysis
This paper challenges the common assumption that higher-order ODE solvers are inherently faster for diffusion probabilistic model (DPM) sampling. It argues that the placement of DPM evaluations, even with first-order methods, can significantly impact sampling accuracy, especially with a low number of neural function evaluations (NFE). The proposed training-free, first-order sampler achieves competitive or superior performance compared to higher-order samplers on standard image generation benchmarks, suggesting a new design angle for accelerating diffusion sampling.
Key Takeaways
- •Challenges the dominance of higher-order ODE solvers for DPM sampling speed.
- •Proposes a novel, training-free, first-order sampler.
- •Demonstrates competitive or superior performance compared to higher-order samplers on image generation benchmarks.
- •Highlights the importance of DPM evaluation placement for sampling accuracy.
Reference
“The proposed sampler consistently improves sample quality under the same NFE budget and can be competitive with, and sometimes outperform, state-of-the-art higher-order samplers.”